Efficient one-sided Jacobi SVD computation on AMD GPU using OpenCL
Efficient one-sided Jacobi SVD computation on AMD GPU using OpenCL
复制标题
使用 OpenCL 在 AMD GPU 上进行高效单侧雅可比 SVD 计算
DOI：
--
复制
发表时间：
2016
期刊：
影响因子：
--
通讯作者：
Dong Wang
中科院分区：
文献类型：
--
作者：
Jianjing An;Dong Wang
奇异值分解（SVD）是数值计算的重要组成部分，广泛应用于生物医学、气象学和量子力学等众多领域。提高SVD算法的速度和精度成为一个重要问题，因此我们使用OpenCL语言在AMD GPU上研究高效的并行SVD算法。近年来，人们提出了许多用于 SVD 硬件计算的方法，但这些方法都受到速度的限制，我们利用 OpenCL 在 AMD 图形处理单元上提出了一种单侧 Jacobi 并行算法。在本文的前面部分，介绍了SVD算法和单侧雅可比算法，并利用环雅可比排序实现了并行计算。接下来我们给出SVD工作流程并在GPU上实现8×8、16×16、32×32、64×64,128×128和256×256矩阵。通过与 MATLAB 和其他论文进行比较，我们的加速比分别约为 3.25× 和 1.24×。
Singular value decomposition (SVD) an important part of numerical calculation, widely used in many areas such as biological medicine, meteorology and quantum mechanics. Improving the speed and accuracy of SVD algorithm becomes an important issue, so we study efficient parallel SVD algorithm on AMD GPU using OpenCL language. In recent years, there are many approaches for SVD hardware computation have been proposed, however, which are limited by speed and we propose an One-sided Jacobi parallel algorithm on AMD Graphics Processing Unit by using OpenCL. In the front part of this paper, SVD algorithm and One-sided Jacobi algorithm are introduced and by using the Ring Jacobi Ordering we achieve our parallelism computation. The next we give SVD workflow and implement 8×8, 16×16, 32×32, 64×64,128×128 and 256×256 matrices on GPU. By comparing with MATLAB and other paper, our speedups are respective approximately 3.25× and 1.24×.